AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex

Amazon Web Services (AWS) has unveiled Strands Harness, an open-source general-purpose AI agent designed to help developers create ready-to-use AI agents for local or cloud deployment. Strands Harness builds upon AWS's Strands Agents, an open-source Python SDK introduced in May 2025.

Developers using Strands Harness provide the model, tools, and instructions, while the model determines how to approach the task and when to utilize those tools. Marc Brooker, VP and distinguished engineer at AWS, explains that Strands Harness is essentially an enhanced version of the existing Strands SDK, providing preconfigured agents with necessary tools and machinery for long-running tasks. Users still need to manage context, persistent conversations, tool integration, and guide the agent's behavior.

Strands Harness includes a range of default tools, such as file, shell, and web tools, along with built-in management for context, memory, persistent sessions, prompt caching, and delegation to other agents. Developers can customize these components according to their needs, including changing the instructions, selecting the model, controlling available tools, and deciding whether to delegate work to other agents.

Most of Strands Harness does not rely on AWS infrastructure. The agent loop, tools, context management, session handling, and delegation are included in the open-source release and run on the machine executing the agent. However, the underlying model access is routed through Amazon Bedrock, AWS's managed service for accessing and running foundation models.

Developers can override this default and use alternative models from providers like Anthropic, OpenAI, or Google, or even run a locally hosted model using Ollama. The choice of model provider does not necessarily alter the underlying model, but different models can impact the agent's behavior.

AWS claims that Strands Harness is more cost-effective and efficient than similar AI agents, such as Claude Code and Codex, which it compares against. In benchmark tests, Strands Harness outperformed Claude Code and Codex by 45% in cost while maintaining comparable accuracy. However, when DeepSeek Harness is included in the comparison, the cost advantage drops to 28%.

AWS attributes this performance gain to its context-management defaults, which include truncating large tool outputs, compacting context, and attempting to recover within the agent loop if context overflows. This results in significant cost savings, as demonstrated by Terminal Bench 2.1, where Strands Harness running Fable 5 cost 77% less than Claude Code, at $56.29 versus $248.05 across 89 trials.